Can I just stay in this moment with you?
Every now and then you find a liminal space where things are quiet and you can catch your breath. This is one of those spaces.
Happy Friday.

It started with this prompt:
“A highly detailed and realistic analog photograph, 85mm lens aesthetic, Leica camera aesthetic, 8K, UHD, vibrant colors, natural lighting, old stock film. A nude woman with blonde hair styled in an undercut sitting on a park bench at night.”
And that’s how we got the image on the left. And it was a hell of a first image. The pose and the facial expression are captivating [although there might be some anomaly with the feet, I can’t quite tell]
I had been working on the “Thin Blue Lines” video and I was starting to become enamored with the character of Officer Catherine Morales, and I wanted to do something with a similar look.
These were the results:











I loved the look and feel of the photos, but I wasn’t happy with the hair. I liked the images where the hair was longer, so I decided to adjust for that:

A highly detailed and realistic analog photograph, 85mm lens aesthetic, Leica camera aesthetic, 8K, UHD, vibrant colors, natural lighting, old stock film. A nude woman with long blonde hair styled in an undercut sitting on a park bench at night.”
That first iteration was closer to what I wanted; there’s something dramatic about the length and the way it interacts with the light. And I have no issue with short hair, but longer hair frames the face in a certain way and that was the look I was going for.
These were the results of a few generations:

















I was starting to feel burnt out at this time. I was trying to make Thin Blue Lines while juggling a few other things – and my mind kept driftintg to the film that I’m working on which I haven’t made any progress on for about a month.
And something about the images was bothering me – I didn’t like the slender bodies – and it wasn’t a matter of aesthetics – it was an emotional reaction. The women felt frail and that wasn’t the space I wanted to get lost in.

This is where things started to feel right for me:
“A highly detailed and realistic analog photograph, 85mm lens aesthetic, Leica camera aesthetic, 8k, UHD, vibrant colors, natural lighting, old stock film. A nude woman with a muscular and very curvy body, long blonde hair styled in an undercut with pink highlights, sitting on a park bench at sunset.”
That body type is very simlar to the Catherine Morales character in Thin Blue lines; it’s a body built like a sports car; hard and sleek and curves in all the right places; and the long hair covering part of the face and the pink tones create a sense of mystery and playfulness.








So we had our girl now. But something still felt off.
The space she was in didn’t feel right. I wanted something more liminal – familiar and strange at the same time.

And…I didn’t really know if this worked when I first tried it:
“A highly detailed and realistic analog photograph, 85mm lens aesthetic, Leica camera aesthetic, 8k, UHD, vibrant colors, natural lighting, old stock film. A nude woman with a muscular and very curvy body, long blonde hair styled in an undercut with pink highlights, sitting under a statute in a park at sunset.”
And there’s a lot of things wrong with that image; it’s a bad composition; it’s a bad pose; her skin is too shiny for the lighting to be correct. It’s a pretty face, but that’s about it.
But you have to do a lot of iterations to see if something is viable – that’s an important aspect of working with AI. Sometimes you get a good result on the first try. Sometimes you have to generate 100 images until you find one good one.















And there’s some good images in there. But, it wasn’t quite what I was going for.
Up to this point I had been using Grok to generate the images, but I decided to try something different, just to see how another model would tackle the prompt.
Perchance
So I switched to Perchance and these were the results:

This is my favorite of the Perchance generations which were done with this model: https://perchance.org/ai-character-generator
While it’s not realistic, I think the stylization is perfect and I like how her musculature contrastrs with the muscular frame of the statue.
And as always, Perchance created some interesting images, and there’s things I like, but nothing quite works:












































































But at the end of the day, Perchance is free and experimentation costs nothing.
Which made me curious about how other models would fare if I was actually paying money to generate images.
Z-IMAGE TURBO
Z-Image Turbo is a model that Atlas Cloud keeps pushing on me, and it only cots $0.01 per image, so let’s see what it can do:

So I actually like this quite a bit, although part of her leg appears to have disappeared and Z-Image thinks I want the leica camera in the shot.
And that’s an easy fix, but I wanted to stick with the same prompt across all models – and frankly if Perchance can figure it out then a paid model should understand what I mean by aesthetic.




Qwen-Image Text-to-Image Plus
Qwen has been around, but I feel like it’s been overshadowed by other models. To generate an image it cots $0.021:

I’d never used Qwen before, so I didn’t know what to expect. I actually like this visual quality a lot – it’s slightly uncanny but also hyperrealistic. I also LOVE the bizarre proportions of the statue.
I could see myself using this for a future project at some point, because I am curious how some of these images would look in motion.




Wan 2.5
I’ve used Wan 2.5 for video before, but I have never tried out its image generation capabilities before. It costs $0.021 to generate an image, which is pretty reasonable.

So when you say “muscular” Wan 2.5 takes that very seriously. And I like it.
It’s a bit like Perchance, if that model was able to render on the more realistic side, while still keeping the stylization that makes it fun. Notice that her hand appears to have a both an underdeveloped finger and a vestigial 6th finger, which doesn’t seem like a thing that should be happening with a pay per use model.




Seedream v4
Seedream is one of my favorite image editors, but I rarely use it for image generation. So I was excited to see how it would handle the prompt. It costs $0.024 to generate an image:

This is by far one of my favorite images to come out of this project. This probably comes the closest to replicating the “old film stock” look that I wanted. But everything about it is well done; the pose; the body details; the facial detail; the overall composition. One thing I really appreciate is how Seedream interpreted “muscular and very curvy,” because I think on Wan and Qwen the breasts are nice, but they don’t look natural due to the amount of muscle mass on display – but in Seedream everything looks realistic without going too far in one direction or the other.




Grok
I wasn’t done playing with Grok though. Because I wanted to use the “extend” feature for video and I just had a feeling something would eventually look right to me.
And then something slipped through that was unexpected:

Because I never thought about playing with the lighting in that way.
And while everything else had kind of generic “fashion” lighting, this felt more real; more raw.
This image had the liminal quality that I wanted. It’s familiar and strange at the same time.
She looks like something we discovered – a bonfire in the darkness.
And that’s what I wanted.
Because when I was making this I was feeling tired and sick and I just wanted a place to rest my head.
A place to exist outside of the noise of the real world.

Leave a Reply